miroslavmorozov

The sole developer and architect of a production AI assistant integrated with WhatsApp and Telegram using OpenAI API, AWS, and async Python backend handled everything from infra to user-facing logic. Built a custom orchestration pipeline and agentic classification system using OpenAI prompts and regex, enabling prioritization and reducing response times below 5 seconds. Delivered full-stack features including human takeover, automated summaries, lead extraction, and CRM sync - delivereded and deployed to real clients.

Engineered a scalable call outcome classification system using OpenAI GPT, achieving 95% coverage. Outperformed third-party solutions in quality and scalability, ensuring seamless client onboarding and increasing conversion rates by 10%. Fine-tuned large language models (LLMs) for classification tasks, improving model accuracy from 75% to 95%. Designed and implemented a comprehensive funnel analytics tool for conversion analysis, identifying lead quality issues through statistical significance testing. This tool secured client commitment to enhance lead quality, resulting in the expansion of the pilot project and an increase in project volumes. Automated transcription and translation processes for client calls, reducing manual workload and accelerating client onboarding to under 30 minutes.

Led planning and task estimation for a 4-person team, conducted quarterly performance reviews, and mentored new team members. Leveraged insights from other products to enhance team efficiency and productivity. Developed and implemented the Mean Dialog Length (MDL) metric to evaluate dialog quality, demonstrating its correlation with user engagement. Optimized response ranking through iterative model improvements, resulting in a 15% increase in user engagement. Designed a custom pipeline for response relevance using cross-encoder RoBerta models, reducing irrelevant responses from 24% to 10% by distilling models into bi-encoders for enhanced efficiency. Created dashboards to monitor model improvements and performance.